Eduardo Vedes (@eduardovedes) header
Eduardo Vedes (@eduardovedes) profile picture

Software Engineer & AI Consultant | Software Architecture, AI Integration & Team Training

Faro, Portugal
Joined September 2009
Photo by Eduardo Vedes (@eduardovedes)Photo by Eduardo Vedes (@eduardovedes)Photo by Eduardo Vedes (@eduardovedes)Photo by Eduardo Vedes (@eduardovedes)Photo by Eduardo Vedes (@eduardovedes)Photo by Eduardo Vedes (@eduardovedes)Photo by Eduardo Vedes (@eduardovedes)Photo by Eduardo Vedes (@eduardovedes)Photo by Eduardo Vedes (@eduardovedes)Photo by Eduardo Vedes (@eduardovedes)
We may have crossed paths here without an introduction. I’m Eduardo 👋 Programming since 1986. Nine years professionally in software, after a career in civil engineering. Now: AI at Mindera, building One Dev Army, and training teams. What are you working on?
11
3
159
EU AI Act Article 50(1) generally requires providers to make direct interaction with AI clear unless obvious in context. Open the bot on mobile. What does a screen reader announce? digital-strategy.ec.europa.e…
2
1
60
Herdr shows which coding agents are working, blocked or idle. My test would be a handover. Can another engineer find the blocked migration, understand the open question and route it to the person who owns the data? herdr.ink/
34
The first AI draft took five minutes. How long until it was ready to merge? Track implementation, review, correction, tool cost, and defects. Keep active work separate from elapsed time: waiting for review can dominate the latter.
2
2
68
Use the result to change one thing: Broad diffs → smaller delegated changes. Repeated misunderstandings → clearer brief or repo guidance. Manual checking bottleneck → better verification setup. Measure again on comparable work.
1
13
A practical AI coding loop: Define the outcome. Provide missing context. Test requirements. Debug with evidence. Review access rules. Coordinate independent work. Measure through review and correction. Try it on one task this week.
11
Before adding another coding agent, identify the dependency. For a CSV import feature, agents can independently map the upload flow, inspect validation rules, and propose acceptance cases. Give each a clear question and a concrete output.
2
1
53
Before parallel implementation, agree on the API contract and error cases. Define file ownership. Use separate branches or worktrees for edits. Remember: worktrees don't isolate shared databases, test services, or external side effects.
1
16
Delegation brief: Task → Inputs → Owned files → Expected output → Verification. After integration, test the complete user flow. Count coordination and review time. Small, tightly coupled changes may be easier to finish in one session.
16
Spot the missing check: GET /invoices/:id 1. Require a logged-in user. 2. Load the invoice by ID. 3. Return it. Alice can read her own invoice. The test passes. What happens if she supplies Bob's invoice ID?
2
37
Authentication identifies the requester. Authorization checks access to the resource. Enforce the application's ownership or tenant policy on the server. Test Alice's allowed invoice, Bob's invoice, unauthenticated access, and explicit privileged-role rules.
1
11
Ask the agent to trace handler → service → database query. Where is access enforced? Which test proves it? Also review unfamiliar APIs, new dependencies, and sensitive logs. A useful review habit: run the same request as a different user.
9
After three unsuccessful AI fixes, improve the evidence. Capture the smallest reproduction, expected vs actual behaviour, relevant logs, versions, and recent changes. Each attempt should teach you something about the cause.
1
42
Example: a CSV header becomes one field. Inspect the actual input and parsed header. A delimiter mismatch, encoding marker, or parser option may explain it. Isolate the parser in a tiny reproduction. Use the evidence to choose the fix.
1
12
After the fix: keep a useful regression test, run the affected checks, and inspect the diff for unrelated edits. If the session starts looping again, ask: what new evidence would change our diagnosis?
6
AI can write code and tests that agree on the same mistake. Requirement: reject an entire CSV if any row is invalid. Implementation: save valid rows, skip bad ones. Test: confirm valid rows were saved. Green tests. Broken requirement.
1
41
Derive the expected result from the requirement: Input: one valid contact + one invalid email. Expected: report the bad row and save zero contacts. For a bug fix, confirm the test fails against the old implementation for the right reason, then passes with the fix.
1
18
Review test changes as carefully as code changes. Watch for removed assertions, skipped tests, and mocks that bypass the behaviour being checked. Ask AI for edge cases. Review the expected outcomes yourself. Agent agreement isn't proof.
16